Real-Virtual Data Fusion and YOLO11-based Segmentation for Automated Deposition Region Recognition in L-DED Repair Process

Reliable recognition of damaged regions is essential for automated L-DED repair, yet conventional pre-process planning often relies on costly 3D scanning and CAD/CAM alignment. We propose an on-machine, single-RGB-camera framework using YOLO11 instance segmentation with real–synthetic hybrid training data generated via physics-based rendering to address labelled data scarcity. Using a 28-image unseen-object (LOTO) test set from a blade-inspired specimen excluded from training, Hybrid-YOLO11m achieved 94.2% IoU and 97.0% F1, outperforming real-only, synthetic-only, and optimized OpenCV baselines under varying illumination and reflections. Furthermore, a bead-width-based offset compensation strategy reduced the effective risk of under-deposition by approximately 96.8%. The proposed approach enables low-cost, streamlined pre-process planning without 3D scanners for practical DED repair.